I am just wondering is there a way to inverse encoding with SDRs?
I learned that HTM has a sensory part and a motor part.
The sensory part is to capture the data from outside and the motor part is to actually react based on the data.
In the HTM school series on YouTube, I see how the HTM will encode the real life data to SDRs but I think it did not mention how to actually use that data or how to decode those SDRs into some certain reactions.
Am I missing anything? Did any papers or video tutorials talk about that?
No, the videos didn’t touch on the subject much. There are only a few places mentioning it.
There’s a few ways to decode SDRs depending on what you are doing and wish to achieve. If you are predicting a time series, which happened to be encoded using Grid Cells. Then it can be trivially decoded using linear models. Object representations, can be decoded using a classifer. And converting SDR into motor commands is a field of active research. Most implementations just use a SP as a auto-correlation learner (but I doubt it is the best approach) like it is used in a HTM based Q-ish learner.
Here’s one that talks about it the SDR classifier, which is used in NuPIC to convert the TM predicted states back into the raw data format for forecasting applications:
Pattern → pattern is not how cortex works.
That is a deep learning thing.
There is a cluster of local functions at the mini-column level (see below) but at a higher level the main function of cortex is pattern completion / recognition. These patterns are both spatial and temporal. A closely related function is to signal novelty with bursting - this feature is signaled to subcortical structures to draw attention to lack of pattern recall.
A significant part of the function of the cortex is to learn these patterns for subsequent recall.
There - that’s all there is.
The subcortical structures are able to use this to do all the cool things cortex does.
It takes some reflection to see how cortex does it, but this is the same way that the forebrain forms patterns to guide motor drivers - pattern completion. (Still spatial and temporal patterns)
How does those driver patterns get into forebrain? Watch a baby babbling patterns into her body. The reward functions are in the subcortical structures, not the cortex.
